How many buildings does South Africa have, and where are they packed tightly enough for fibre to pay off?
For Day 16 of the 30 Day Map Challenge, the theme is Collaborative. So instead of building a map on my own, I built one with my colleague James Dales, using the same data in two different tools. James mapped it in Icon Map for Fabric; I mapped it in Power BI with Icon Map Slicer.
The subject comes from a gap in the series. After fifteen days I had covered mining, energy, conservation and marketing, but not telecoms or last-mile logistics. Both industries ask the same question: where are buildings close enough together for a network, or a delivery round, to pay for itself?
The answer starts with a number: South Africa has about 33.1 million building footprints. Together they cover about 2,580 km², roughly 0.2% of the country.
Explore the interactive report
The report opens on a national 3D view. Each column is a 4.5 km geohash cell:
- Height is the number of buildings in the cell, on a square-root scale so that small towns stay visible next to larger city metropoles.
- Colour is the cell's average buildings per km²: under 10, 10 to 49, 50 to 199, 200 to 799, and 800 and over.
Zoom in and the map changes scale. From zoom 8 the columns give way to 1 km cells, coloured by deployment zone. From zoom 11 come 150 m cells, and from zoom 12 every building footprint:
- Fiber: 200 or more buildings per km²
- Fixed wireless: 50 to 199
- Scattered: under 50
The Show on map slicer switches the zoomed-in layers between buildings and the 100 m population grid. In population mode, squares rise 10 m per person.
On the right, four cards, two zone charts and a municipality table follow the map. Pan or zoom and they recount what's in view, down to street level. A Province > Municipality slicer filters the whole page. The table lists the municipalities in view, with their whole-municipality figures.
Interactive report: Day 16, Collaborative
What the map doesn't show: actual fibre coverage. The zones are indicative, based on building density alone. They say nothing about distance to an existing network, terrain, take-up or cost. And a building isn't a household: a block of flats is one footprint.
Starting with 33 million buildings
The buildings come from the Overture Maps Foundation buildings theme, release 2026-10-21.0. Overture merges footprints from OpenStreetMap, Microsoft and Google into one dataset under the ODbL. That is a collaboration in itself: three very different contributors, one schema.
James read Overture's cloud storage directly with DuckDB and clipped it to South Africa: 40.1 million buildings in the bounding box, 33,082,488 inside the country. Each footprint was reduced to its centre point for counting.
The footprints are machine-detected, so some buildings are missing and some aren't buildings. Detection is also uneven. A large rectangle of the Karoo has sparser Google footprints, which pushes towns such as Colesberg and Beaufort West below the thresholds. You can see it on the national map: a straight-edged square under the “South Africa” label, from about 22°E to 26.5°E and 30°S to 33.5°S (roughly 420 × 390 km, taking in Beaufort West, De Aar, Colesberg and Graaff-Reinet). Inside it, only 30–45% of the 4.5 km cells contain any building; on every side, 90% or more do. Settlement doesn’t stop along lines of longitude, so this is one of Overture’s sources ending at a tile boundary, not an empty Karoo. Treat the counts as estimates, not a register.
Adding 61.7 million people
Population comes from GHS-POP R2023A, the European Commission Joint Research Centre's 100 m grid for 2025: about 11.8 million cells and 61.67 million people in South Africa.
The grid tells a different story from the buildings. Each 100 m square's people were shared equally among the buildings inside it. But about 10% of people live in squares with no detected building at all. So every people figure in both reports comes from the population grid, not from the buildings.
One lakehouse, one set of definitions
The collaboration ran through a shared Fabric lakehouse. James built the tables and tiles there. I copied what I needed into my project:
- a geohash-5 table: 50,970 cells of about 4.5 km
- a geohash 6 and 7 table: 666,843 cells of about 1 km and 5.6 million of about 150 m, each with its centre, buildings, people and zone
- 212 municipalities, using the OpenStreetMap outlines from Day 11
- three PMTiles files on blob storage: the 1 km and 150 m cells (243 MB), every building footprint (1.6 GB) and the 6.1 million populated 100 m squares (347 MB)
We agreed one set of definitions before either of us built anything. Zones are set on the 1 km cells, using each cell's real area. Smaller cells and every building take the zone of the 1 km cell they sit in. Because both reports read the same tables, the national totals agree exactly: 33,082,488 buildings and 61.67 million people at every level.
Geohash: cells you can drill by prefix
Geohash is new to this series (I'm keeping hexagons for Day 20). A geohash is a short code for a rectangle of latitude and longitude, and each extra character splits the cell into 32. At South Africa's latitudes:
- 5 characters (
kdcsk) is about 4.5 × 4.5 km - 6 characters (
kdcskn) is about 1.1 × 0.6 km - 7 characters is about 150 m
Cells nest by prefix, so moving from one level to the next needs no spatial join. Delivery apps and telecoms use geohash for exactly this kind of zoning.
The cells are rectangles in latitude and longitude, so they cover a little less ground towards the poles. Across South Africa the effect is only a few percent, and every density uses the cell's real area.
Icon Map Slicer reads geohash codes natively. The geohash-5 layer is a cells layer with its source set to Geohash IDs, the code in the ID field and the column height in the weight field. There's no geometry in the model at all.
One catch: rows drawn from geohash IDs must leave latitude and longitude empty. If a row carries coordinates as well, the visual treats it as a point and skips it.
Keeping the figures alive at street level
Icon Map Slicer can filter the rest of the page by the map's field of view: the visuals recount whatever has a point inside the visible area. My first version filtered on the geohash-5 cell centres. That worked nationally, then broke completely at street level. Zoomed in on a small settlement, the view was only about 2 km across. Geohash-5 centres are 4.5 km apart, so often none was on screen, and every card showed (Blank).
The fix was to move the view filter to the geohash-6 cell centres, about 1 km apart. The cards and zone charts now count those 1 km cells, so they work down to a single neighbourhood.
The model links municipalities, geohash 6 and geohash 5 in one chain. That way the slicer, the lasso, a click on a column and the map view all reach every visual.
Tile layers the Format pane can't build
The street-level layers are vector tiles, and that's where the Power BI report needed help from outside the visual's format pane.
Icon Map Slicer has catalog layer slots that draw PMTiles with their own styling: fill rules, extrusion, outlines and zoom ranges. Each slot stores its whole configuration as one JSON object in the visual's settings. The Format pane builds that object only when you add a layer from the Icon Map Catalog. There is no dialog for "my own PMTiles file, styled by this field, with these rules".
So I wrote those configurations straight into the report's PBIP files, in the map's visual.json. That's possible because a PBIP report is a folder of readable JSON. There are four layers:
- 1 km cells from zoom 8 to 11 and 150 m cells from zoom 11, filled by each feature's
zonevalue - every building footprint from zoom 12, filled by zone, with a thin dark outline so teal buildings stand out on teal cells
- 100 m population squares, extruded by height field
The Show on map slicer then switches them on and off. Each layer's visibility reads a yes/no column from a two-row selector table, so choosing Buildings or Population shows one group of layers and hides the other.
A legend-only layer
The same trick fixed the legend. The geohash-5 columns are coloured row by row from a hex column in the data. A layer coloured that way can't list its categories in the map's legend, so my first version used a text box.
A catalog layer's legend, on the other hand, comes from its named fill rules. So I added a fifth catalog layer that draws nothing:
- it has five rules, under 10 to 800 and over, each in the exact column colour, on a field that doesn't exist, at zero opacity
- it points at the population tiles, which only start at zoom 8, so below zoom 8 it never even downloads a tile
The map's How dense? legend now shows the geohash-5 bands at national scale and the zones once you zoom in. Again, this is only possible by writing the layer's JSON into the PBIP.
Keeping it rebuildable
Because some settings can only live in the JSON, the whole report is generated from a spec, and a short post-processing script adds what the spec can't express. When I changed something in Power BI Desktop, such as the hierarchy slicer or the card spacing, I copied it back into the spec, so a rebuild never undoes a manual edit.
Design choices
- Two colour schemes, deliberately. The columns show average density in five bands. The zoomed-in layers and both bar charts show the three zones (teal for Fiber, amber for Fixed wireless, pink for Scattered). They share the 50 and 200 cut-offs, which is why they look like one scheme. More on that below.
- A darker top band. James uses near-white for 800+ on a dark basemap. On the light DataViz basemap that colour had a contrast of 1.05:1, so it's deep teal here.
- Height on a square-root scale. On a linear scale, Johannesburg and Durban would flatten everything else.
- A lighter 150 m wash. At 35% opacity, the footprints read through the cells.
- Whole-municipality figures, labelled as such. The table's figures are bigger than the cards' because they cover whole municipalities, and the title says so. It has no total row, because adding up percentages across municipalities would mislead.
What the map reveals
Most of South Africa is already dense enough for fibre. 79% of buildings (26.2 million) and 86% of people (52.8 million) are in 1 km cells with 200 or more buildings per km². Gauteng is one continuous block from Soweto to Pretoria.
Fixed wireless is a large middle ground. About 4.6 million buildings and 6.2 million people are in the 50 to 199 zone. That's about 1.3 million buildings each in KwaZulu-Natal, the Eastern Cape and the edges of Gauteng.
The hard-to-reach population is rural, not small. About 2.3 million buildings and 2.6 million people are in scattered zones:
| Province | People in scattered zones | Share of its people |
|---|---|---|
| Eastern Cape | about 464,000 | 6.9% |
| Western Cape | about 381,000 | 4.8% |
| KwaZulu-Natal | about 348,000 | 3.2% |
| Free State | about 321,000 | 11.8% |
| Northern Cape | about 190,000 | 13.5% |
| Gauteng | about 174,000 | 1.0% |
The biggest numbers are in the Eastern Cape, the Western Cape and KwaZulu-Natal. The highest shares are in the Northern Cape and the Free State, partly because of the thinner Karoo footprints. The former homelands of the Eastern Cape and KwaZulu-Natal, and the Western Cape's farmland, stand out in the 3D view.
A 4.5 km average hides the towns. A geohash-5 column on the edge of Kimberley has 3,127 buildings over 21 km², so 149 per km² and coloured amber. But 76% of those buildings sit in five 1 km cells dense enough for fibre; the rest is mostly empty land. Across the country, 70% of the buildings in amber columns are actually in fibre zones. The column colour answers "how built-up is this square?" The zone answers "how dense is the neighbourhood each building is in?" Only the second is the deployment question.
The thresholds matter too. James found that at 100 buildings per km², 92% of people would be in fibre territory; at 400, it's 73%. None of this proves fibre would pay. It shows where density alone makes it plausible.
What I learned from Day 16
1. Agree the definitions before the tools. Because James and I agreed thresholds, cell levels and sources first, two reports in two products give the same national totals to the last building.
2. The view filter needs points finer than the view. A map-view filter on 4.5 km cell centres fails at 2 km of zoom. Filter on something finer than the closest view you expect.
3. PBIP opens doors the Format pane doesn't. Self-hosted PMTiles with field-driven styling, and a legend-only layer, exist only in the report's JSON. Treating the PBIP as source code made them possible, and keeping a spec made them repeatable.
4. An average can look like a classification. When two colour schemes share cut-offs, say plainly which one a colour belongs to.
Final thoughts
The Day 16 theme was Collaborative, and the collaboration ran at three levels.
- The data: Overture's footprints are themselves a collaboration between OpenStreetMap volunteers, Microsoft and Google.
- The pipeline: a shared lakehouse with one set of definitions.
- The maps: two people building in two tools, Icon Map for Fabric and Power BI.
For anyone planning networks, delivery zones or service areas, the reusable idea is the geohash ladder. Count everything into nested cells, decide on the finest level that still has meaning (here, 1 km), and let the map move between levels by prefix. The arithmetic stays simple, and the decision sits at the level where it actually belongs.
Sources and attribution
- Buildings: Overture Maps Foundation, buildings theme, release 2026-10-21.0, from OpenStreetMap contributors, Microsoft and Google. Licence: ODbL. https://overturemaps.org. Extracted and clipped to South Africa by James Dales, November 2026.
- Population: Schiavina, M., Freire, S., Carioli, A., MacManus, K. (2023). GHS-POP R2023A: GHS population grid multitemporal (1975–2030), epoch 2025, 100 m. European Commission, Joint Research Centre. Licence: CC BY 4.0. DOI: 10.2905/2FF68A52-5B5B-4A22-8F40-C41DA8332CFE.
- Municipal outlines: © OpenStreetMap contributors (the 212 municipalities used on Day 11). Licence: ODbL.
- Geohash cells, zones and PMTiles: built in Microsoft Fabric for the Icon Map for Fabric version of this map.
Methodology note: Each Overture building footprint was reduced to its centre point and counted into geohash cells at precisions 5, 6 and 7. GHS-POP 2025 people were summed into the same cells from the 100 m grid, including cells with people but no detected buildings, so people total 61.67 million at every level. Buildings per km² use each cell's real area. Zones are set on geohash-6 cells (Fiber 200+, Fixed wireless 50–199, Scattered under 50) and inherited by geohash-7 cells and individual buildings. Geohash-5 column colour is the cell's own average density, not its zone. Municipality is assigned by cell centre. In the Power BI report, cards and charts count geohash-6 cells whose centre falls in the map view, so totals at the edge of the view are rounded to whole 1 km cells. Building counts are machine-detected estimates, not a register, and a building is not a household. Zones are indicative of density only, not a commercial feasibility assessment.